Cuttings Accumulations Prediction in Deviated & Horizontal Wells with Dimensionless Data-Driven Models

Abstract Effective cuttings removal in deviated and horizontal wells is essential for improving drilling efficiency and preventing non-productive time (NPT) caused by hole-cleaning issues. While various numerical models have been developed to simulate cuttings accumulation in wellbores, only a subset of these models can be employed for real-time operations due to their complexity and lengthy computational requirements. This paper compares the performance of various data-driven (machine learning) models in monitoring cuttings bed accumulation in real-time during drilling operations. The construction of these data-driven models relies on the analysis of hundreds of bed height measurements obtained from ten flow loops. These models incorporate unique dimensionless parameters and are trained on a diverse dataset encompassing a wide range of drilling conditions. These conditions include variables such as the rate of penetration (ROP), drilling flow rate, drillstring rotation, hole eccentricity, wellbore hydraulic diameter and inclination, drilling fluid rheological parameters, and cuttings (solid) density and size. Five different data-driven models are evaluated: linear regressor (LR), deep neural networks (DNN), support vector regressor (SVR), random forests (RF), and extreme gradient boosting regressor (XGBoost) algorithms. Additionally, the performance of the developed models is assessed against previously unseen datasets to ensure fair evaluation. Comparisons are also made with the Duan correlation (a mechanistic model) to evaluate the accuracy and limitations of the data-driven models. A total of ten dimensionless parameters are devised to estimate bed height accumulation using the Buckingham-Π theorem and Pearson correlation. The results indicate that both the RF and XGBoost models exhibit accurate estimations of bed thickness, achieving root mean square error (RMSE) and mean absolute percentage error (MAPE) values around 0.07 and 13%, respectively. Furthermore, these two models demonstrate strong generalization capabilities and precision in estimating bed thickness, with a MAPE below 20% when validated against unseen datasets and compared to the Duan model. In contrast, the DNN model is observed to be less accurate than the RF and XGBoost models, though a majority of its predicted points still fall within the ±20% tolerance envelope. On the other hand, both the SVR and LR models exhibit poor accuracy in capturing the underlying relationship between input parameters and the target variable, as evidenced by their scattered residual values. Utilizing the Shapley additive explanations (SHAP) approach and RF feature analysis, the study identifies the Froude number as having high feature importance while negatively impacting bed height predictions. Conversely, the inlet feed concentration and annular eccentricity significantly positively contribute to bed height prediction. In conclusion, the data-driven (machine learning) models developed in this study offer a reliable means of real-time prediction for cuttings bed thickness during drilling operations. By eliminating the need for complex numerical models with extended computational times, these models empower proactive decision-making, thus enhancing drilling efficiency and minimizing NPT resulting from inadequate hole cleaning.